Scattering of Ocean Surfaces in Microwave Remote Sensing by Numerical Solutions of Maxwell Equations. Tai Qiao

Transcription

1 Scattering of Ocean Surfaces in Microwave Remote Sensing by Numerical Solutions of Maxwell Equations by Tai Qiao A dissertation submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy (Electrical Engineering) in the University of Michigan 2018 Doctoral Committee: Professor Leung Tsang, Chair Professor Anthony Grbic Professor Christopher Ruf Professor Fawwaz T. Ulaby

2 Tai Qiao ORCID id: Tai Qiao 2018

3 DEDICATION To my family, With love and gratitude. ii

4 ACKNOWLEDGEMENTS I would like to express my thanks to my PhD advisor Professor Leung Tsang, for my gratitude to him to share his knowledge with me. During the four years PhD study, I have learned so much from him. One thing I have learned from him is his attitude towards research. He spends most of his time during the day on research. For him, the truth of science always comes first. Even on weekends and deep into the night, we sometimes receive his s or calls about new ideas about research. This touched me so much especially considering the fact that he is about to retire. Another thing that impressed me about him is his ability to broaden research topics and his concentration on research. For a particular problem, he has always been able to point out the key point and propose new ideas to draw other people s attention. He treats everything carefully. He sometimes talks about his personal life to us. His way of carefully handling everything and fulfilling his duty perfectly have influenced me very much. One particular thing that I want to mention is that he is so good at teaching. Even for a difficult concept like electromagnetics, he is able to find a way to interpret to students easily. I really enjoy the time during his class, during which I am so focused. This has never happened to me ever since my undergraduate study. Also, I would like to thank my other 3 PhD committee members: Professor Fawwaz T. Ulaby, Professor Christopher Ruf, and Professor Anthony Grbic. Professor Ualby, Professor Ruf and Professor Grbic are all world famous scientists in their own areas. It is my great pleasure to have them on my committee. iii

5 During my PhD study one of my major source of support was the National Aeronautics and Space Administration (NASA). For the study of ocean scattering problems, I have been able to collaborate with Dr. Simon Yueh from the Jet Propulsion Laboratory (JPL) and Dr. Douglas Vandemark from the University of New Hampshire, Frederic Nouguier and Bertrand Chapron from the Laboratoire d'océanographie Spatiale, Institut Francais de Recherche pour lâ'exploitation de la Mer in France. Without the help from distinguished scientists, I would not have finished. The first 4 months of my PhD study was spent at the University of Washington in Seattle. Here I also want express my thanks to my friends, staff and professors back at the University of Washington. I have also received many help from my group mates and lab mates. They are Wenmo Chang, Shaowu Huang, Xiaolan Xu, Shurun Tan, Huanting Huang, Jiyue Zhu, Mohammadreza Sanamzadeh, Weihui Zhu, Maryam Salim, Yanlei Du, Liujun Zhu. and for those RadLab-mates, just to name a few, Dr. Jiangfeng Wu, Dr. Yang Liu, Dr. Adib Nashashibi, Dr. Leland Pierce, Mostafa Zaky, Mohammad Mousavi, Mani Kashanianfard, Zhanni Wu, Xiuzhang Cai, Han Guo, Weitian Sheng, and Nikolaos Chiotellis. Many thanks will also go to the all the staff and professors in the Radiation Lab. During my PhD study, I have been a GSI for Professor Brain Gilchrist. I have really learned so much interesting knowledge about link design which is quite different from my research. Most of my research work is related to high performance computing. Flux facility at the University of Michigan and Stampede at XSESE are the major two computational resources for me. As I approach completion of my PhD study, I want to express my thanks to the staff at these two supercomputing centers. iv

6 And of course the people who I have spent most of life time with are my family: my father Shungao Qiao, my mother Shuping Jiang, and my older sister Yong Qiao. Without their love and support I could not finish this thesis. Also I want to thank my paternal grandparents who are now both over 80. Now they live peacefully in a small countryside home back in my hometown in Northern China. Most of the past 4 years time were spent at the University of Michigan, Ann Arbor. I really like the university and the city of Ann Arbor. I have many good memories in different places on the campus and the city of Ann Arbor. I wish one day I can come back to the university and work in it. Tai Qiao April, 2018 Ann Arbor v

7 TABLE OF CONTENTS DEDICATION ACKNOWLEDGEMENTS LIST OF FIGURES LIST OF TABLES ABSTRACT ii iii ix xii xiii Introduction Surface scattering of sea surfaces Scattering of anisotropic ocean spectrum Neighborhood impedance boundary condition (NIBC) PBTG in conjunction with SMCG Overview of the thesis 5 Sea Surface Radar Scattering at L-Band Based on Numerical Solution of Maxwell s Equations in Three Dimensions (NMM3D) Introduction Ocean spectrum and surface generation PMCHWT surface integral equation and numerical implementation using RWG and SMCG Ocean surface profile generations and L-band backscattering variations with kl and ku NMM3D results and discussions Conclusions 25 Appendix A 27 Appendix B 31 Radar Scattering of Ocean Surfaces with Anisotropic Ocean Spectrum using NMM3D simulations 33 vi

8 3.1 Introduction Ocean spectrum Nonlinear ocean surface profiles A newly developed directional wave spectrum Results and discussions Scattering of nonlinear Elfouhaily surfaces Scattering of Apel spectra surface with modified angular spreading function Conclusions 45 Scattering of Lossy Dielectric Surfaces in Full Wave Simulation of Maxwell s Equations with Dense Grid and Neighborhood Impedance Boundary Conditions Introduction Formulations Dual surface integral equation Impedance boundary condition Tapered incidence wave Pierson-Moskowitz spectrum Neighborhood impedance boundary condition Results and Discussions Comparison between IBC and benchmark solution Single dense grid and single coarse grid Slow convergence of dense grid NIBC for alleviating numerical issues brought by dense grid Conclusions 64 Fast Computational Method of Combing Low Rank Property and SMCG for Scattering of Ocean Surfaces Introduction Formulation Results Conclusion 77 vii

9 Conclusions 78 BIBLIOGRAPHY 80 viii

10 LIST OF FIGURES Figure 1: Isotropic Durden-Vesecky surface elevation spectrum (solid lines) for wind speeds 5 m/s, 8 m/s, and 10 m/s. Magenta and blue lines denote lower and caupper limit. The lower limit k l corresponds to a surface size of Upper limit k u corresponds to x y 1/ 32. Red and green lines denote the incident microwave wavenumber and Bragg wavenumber respectively. 11 Figure 2: Comparison of L-band NMM3D scattering amplitude distributions for 300 realizations for co-polarized cases VV (left), HH(right) and for wind speed of 5 m/s using the isotropic DV spectrum and an incidence angle 29 o. The red curve indicates that for a Rayleigh-distributed process. 20 Figure 3: In-plane bistatic VV HH for wind = 5m/s isotropic DV spectrum 21 Figure 4: L-band backscatter at wind speed = 5 m/s for different incidence angles: 29, 39, 46 degrees and isotropic DV spectrum compared with upwind Aquarius data. 21 Figure 5: L-band backscatter for wind speed =10 m/s and incidence angle= 46. Upper two panels are VV and HH, and bottom is for VV/HH. TSM, NMM3D, SPM results use the anisotropic DV spectrum. 23 Figure 6: Comparison of NMM3D and two scale model of VV/HH with Aquarius at 29 o, 39 o, and 46 o. NMM3D is using isotropic DV spectrum for 8 m/s and anisotropic DV spectrum for 10 m/s. TSM is derived using the anisotropic DV spectrum for wind = 8 and 10m/s. Each data point is obtained by averaging over all wind directions. 24 Figure 7: Polarization ratio for wind speed 10m/s and 26 incidence. The horizontal axis is different cutoff wavenumbers. Vertical axis is VV/HH ratio in db. Each data point is obtained by averaging over all wind directions. 24 Figure 8: SMCG with centroid as integration points 29 Figure 9: SMCG in conjunction with RWG basis functions 30 Figure 10: Geophysical Model Function of Aquarius scatterometer data versus wind direction of beam 1 at L band. Top figure is VV, middle figure is HH, bottom figure is VH. The four curves correspond to wind speeds 5, 8, 10, 15 m/s. 35 ix

11 Figure 11: Comparison of Ocean (Pierson-Moskowitz), Gaussian and Exponential Spectrum for the same RMS height. Correlation length for Gaussian and exponential spectrum density is 10 times rms height. 36 Figure 12: Computer generated linear ocean surface at 5 m/s using Elfouhaily spectrum. The dimension is 15.2m by 15.2m. 37 Figure 13: Extracted ocean spectrum for 5m/s with and without horseshose patterns of Choppy Wave Model for upwind and downwind. 38 Figure 14: (k) of Du s spectrum for wind speed 5 8 and 10 m/s. 40 Figure 15: Comparison of upwind and crosswind spectrum with the new spreading function by Du and Apel spectrum for 5m/s. Left figure is the zooming in of the right figure for k corresponds to 10 rad/m to 50 rad/m. 41 Figure 16: Incoherent bistatic scattering coefficients by NMM3D for 5 m/s at L band using Apel spectrum and modified angular spreading function. Incidence angle is 38 degree. 43 Figure 17: Comparison of backscatters of AIEM, NMM3D, Aquarius at L band for VV, HH and VH for different wind directions. First row: 29 incidence; Second row: 39 incidence; Third row: 46 incidence. The columns are VV HH and VH returns, respectively. 44 Figure 18: Polarization ratio comparison between NMM3D, AIEM, Aquarius for upwind for the 3 beams: 29, 39, 46 at wind speed 5m/s. 45 Figure 19: Amplitude of the Pierson-Moskowitz spectrum for along-wind direction and wind speeds U19.5=4, 8 and 10 m/s 54 Figure 20: 3D Green function comparison of air and sea water as a function of propagating 55 Figure 21: Fast method for band matrix inversion 56 Figure 22: Surface fields comparisons between IBC and benchmark solution for TE incidence at 5m/s. 58 Figure 23: Surface fields comparisons between IBC and benchmark solution for TM incidence at 5m/s 59 Figure 24: Comparison of surface currents for SDG and SCG. 60 Figure 25: Residual error changing with iterations for two kinds of surfaces: 30λ long surface with 60 points/ λ (blue curve) and 90 λ long surface with 20 points/ λ (red curve) for both TE (left figure) and TM incidence (right figure). 61 Figure 26: Comparison of matrix M between NIBC and IBC with benchmark method (top) ocean surface (bottom) soil surface 63 x

12 Figure 27: Surface fields errors of IBC and NIBC compared to benchmark solution (top) ocean surface (bottom) soil surface 63 Figure 28: Condition number comparison for IBC, NIBC, and benchmark method for ocean surfaces 64 Figure 29: Rank comparison for rough surface and flat surface. Source patch is 16th patch and field patch is from 1 to Figure 30: Steepest descent path and Sommerfeld integration path (SIP) 69 Figure 31: Error of green function calculated by steepest descent method compared to truth value 75 Figure 32: Surface fields comparison between SMCG, PBTG-SDM-FFT and brute force solution. Top figure is for u comparison, bottom figure is for ψ comparison. 76 xi

13 LIST OF TABLES Table I: Upper and lower spectral limits for varying sampling point densities and surface sizes12 Table II: Rms for isotropic DV spectrum at different wind speed in terms of longest surface wavelength 16 Table III: VV and HH for varying discretizing densities at wind = 5 m/s with 39 incidence angle and isotropic DV spectrum compared with upwind Aquarius data 17 Table IV: VV and HH for L band 39 incidence isotropic DV spectrum compared with Upwind Aquarius data 18 Table V: VV and HH of surface w/o horseshoe using NMM3D 43 Table VI: A comparison between PEC, IBC and dual SIE 48 xii

14 ABSTRACT The purpose of this thesis is to study the nature of electromagnetic scattering by the sea surface using numerical methods. Sea-surface scattering has long been studied using various analytical methods. These analytical methods include the two scale method (TSM), the small-slope approximation (SSA), the small-perturbation method (SPM), the Advanced Integral Equation Method (AIEM), and the Geometrical/Physical Optics (GO/PO) method. These analytical methods rely on making approximations and assumptions in the modelling process. Some of these assumptions undermine their applicability in a wide range of situations. The input for analytical methods are usually the ocean spectrum. In real implementations, there are 2 sources of uncertainty in such approaches: (1) the analytical methods have a limited range of applicability to the surface scattering problem; the approximations made in these methods are questionable and (2) the various ocean spectra are another source of uncertainty. We earlier applied a numerical method in 3-dimensions (NMM3D) to the scattering problem of soil surfaces. Through comparison with measured data, we established the accuracy and applicability of NMM3D. We see a drastic increase of ocean remote sensing applications in recent years. It is thus feasible to extend NMM3D to the sea-surface scattering problem. Compared to soil, sea water has a much higher permittivity, e.g., 75+61i at L-band. The large permittivity dictates the need for using a much denser mesh for the sea surface. In addition, the root mean xiii

15 square (rms) height of the sea surface is large under moderate to high ocean wind speeds, which requires a large simulation area to account for the influence of long scale wave like gravity waves. Compared to the two-scale model commonly used for the ocean scattering problem, NMM3D does not need an ad-hoc split wavenumber in the ocean spectrum. Combined with a fast computational algorithm, it was shown that NMM3D can produce accurate results compared to measured data like the Aquarius missions. TSM could also match well with Aquarius provided with a pre-selected splitting wavenumber. But it was observed that the result of TSM changes with different splitting wavenumbers. It is seen that TSM is fairly heuristic while NMM3D can serve as an exact method for the scattering problem. On the other hand, through our study of NMM3D, we found that with a fine grid, the final impedance matrix converges slowly and also it becomes hard to perform simulations for a large surface. This has provoked us to (1) solve low convergence problem for a dense mesh and (2) resolve difficulties in simulations of large surfaces. Inspired by the existing impedance boundary condition (IBC) method, we proposed a neighborhood impedance boundary condition (NIBC) method to solve the slow convergence problem caused by the dense grid. Different from IBC where the surface electric field and the surface magnetic field are related locally, NIBC relates the surface electric field to the magnetic field within a preselected bandwidth BW. Through numerical simulations, we found that the condition number can be reduced using NIBC. Errors of NIBC are controllable through changing BW. We applied NIBC to various wind speeds and surface types and found NIBC to be quite accurate when surface currents only suffer an error norm of less than 1%. On the other hand, in contrast to the large permittivity of sea water, the air permittivity is small and has a rather smooth green function. Utilizing the Steepest Descent Method (SDM), we xiv

16 successfully combined the Fast-Fourier-Transform (FFT) with the Physics-Based Two Grid (PBTG) method. PBTG states that a dense grid is needed for a medium with a high dielectric constant while it is not necessary for air. Eventually, our analysis led to a dense grid, e.g., 64 points per wavelength for green function in the lower medium and a coarse grid, e.g., 8 points per wavelength for green function in air. Utilizing NMM3D, the anisotropic property of ocean backscatter for different wind direction is studied using two different ocean spectra. It is concluded that at L band, the dominant scattering mechanism is Bragg scattering. xv

17 Introduction This thesis is dedicated to the study of microwave scattering by the sea surface using numerical methods. Usually, sea-surface scattering is modeled by analytical methods like the Advanced Integral Equation Method (AIEM), the Small Slope Approximation (SSA), and the Two Scale Model (TSM), among others. These analytical models involve the use of approximations in the modeling process, and thus have a limited range of implementation. In contrast, the numerical method makes no approximations in the modeling process. Surface integral equations are derived from Maxwell equations directly and then solved by the moment method. The input to NMM3D is the real surface profile generated from the ocean spectrum. Unlike a soil surface, the sea surface has a large permittivity and across a large rms height. These two features of sea water and sea surface requires a fine grid up to 64 points per wavelength a large simulation area. A dense grid results in a slow convergence rate for the final impedance matrix equation and a large surface will bring the CPU and memory burdens for the computing clusters. The rest of the thesis will try to solve these problems. 1.1 Surface scattering of sea surfaces Derived from Maxwell equations, the surface integral equations (SIE) are the governing equations for scattering problems. The unknowns are the surface electric fields and surface magnetic fields. In solving the surface integral equations using moment methods, Rao-Wilton- Glisson (RWG) functions have been used as the basis and testing functions. The ocean surface 1

18 profile is generated using linear method with an ocean spectrum. The used ocean spectrum is truncated with a lower limit and upper limit. The lower limit and upper limit are related to the surface length and surface discretization density. Then the surface profile is meshed into triangles where RWG functions can be applied. In solving the final impedance matrix equation, the most computationally intensive part is the product of the matrix by a column vector. To expedite this process, we have implemented Sparse Matrix Canonical Grid (SMCG) method. This method is based on the Taylor expansion of Green function about the flat surface. We project the green function onto the flat surface, then a 2 dimensional fast fourier transform can be carried out. We also combine the RWG basis function with the SMCG method to achieve both accurate and fast computation. Since we are simulating a finite surface, tapered incidence wave have been used to avoid edge diffraction. As an approximation to the relation between surface electric field and surface magnetic field, impedance boundary condition (IBC) has been used widely to surface scattering problem. Out findings are that IBC works fine for surface with a large radius of curvature and large permittivity. But for ocean surface, the capillary waves are the main contributor to backscattering. IBC is not so applicable. We compared our NMM3D results with other analytical methods and also Aquarius measurement data. It is showing that both NMM3D and TSM can predict well matched backscatter with Aquarius. But TSM needs a pre-selected splitting wavenumber to dive the ocean spectrum into SPM part and GO/PO part. When we change the splitting wavenumber, the scattering coefficients produced by TSM will also change. The standard of selecting the splitting wavenumber is like data fitting to produce the best matching with data. Thus TSM is pretty heuristic and cannot be regarded as an exact solution to electromagnetic scattering problem. 2

19 1.2 Scattering of anisotropic ocean spectrum At L band, a new puzzle observed from recent radar missions appeared. As a function of azimuthal dependence, upwind radar scattering cross sections at L band are generally observed to be larger than cross wind radar cross sections. However, at lower winds near 3-5 m/sec, a reversal has been observed where the cross wind radar cross section exceeds upwind. This phenomenon is labelled as a Negative Upwind-Crosswind (NUC) asymmetry. In this thesis we also try to find the reasons lead to NUC by NMM3D. Two candidate anisotropic spectra are explored in Chapter 3: 1) the Elfouhaily Spectrum with and without an imposed nonlinear horseshoe pattern; 2) the Apel Spectrum with an empirically-modified spreading function. The modified angular spreading function is obtained using the advanced integral equation EM scattering method (AIEM) and satellite measurements and where millimeter and centimeter waves crosswind are allowed to be stronger than that of the upwind/downwind. Simulation results show that horseshoe pattern do not cause NUC to happen while the new angular spreading function can reproduce NUC. This leads to the conclusion that at L band Bragg scattering dominate the backscattering. 1.3 Neighborhood impedance boundary condition (NIBC) Through doing 3D scattering problem of sea surface using numerical method, some numerical problems arose. The sea water permittivity is very large compared to air. The wavelength in sea water is 8 times shorter than that in air at L band. This requests a dense mesh when preforming moment method to the surface of sea surface. However, a dense grid will create some significant numerical problem including slow convergence, large CPU and memory consumption. Inspired by the usual Impedance Boundary Condition (IBC), we are thinking of relating the surface electric field with the surface magnetic field accurately. IBC relates local electric field 3

20 with local magnetic field. This means the local electric field is only related to the magnetic field at the same spot. This approximation is made for flat surface with an infinite permittivity. But the truth is surface electric field can be incited by anywhere resided magnetic field. But thanks to the highly lossy feature of sea water, local electric field are only influenced by a small range of magnetic fields that surround it. Thus we can use a band matrix to represent such a relationship. We are testing the NIBC in 2D scattering problem. Results show that 1) the error of NIBC is controllable by changing the bandwidth we are using. 2) NIBC is applicable to a wind range of ocean winds conditions and various surface types. Compared to IBC, NIBC is more accurate in that the surface fields only have an error norm of less than 1%. More importantly, it is found that the resulted equation has a much condition number than the usual dual surface integral equation. 1.4 PBTG in conjunction with SMCG For scattering of random rough surfaces, we have kept mentioning sparse matrix canonical grid (SMCG) as our fast method in moment method for accelerating matrix vector multiplication. And also dense grid should be used due to the large permittivity of sea water. But on the other hand, the permittivity of air is relatively small. So there is a need to use dense grid for green function in the sea water but not for the green function in the air. The goal of Chapter 5 is to achieve that dense grid used in sea water but coarse grid in the air. Similar concept has been proposed by Tsang during the 1990s and 2000s and is called Physics-Based Two-Grid Method (PBTG). With the help of steepest descent method (SDM), a new equation is formulated and through which we have been able to combine Fast Fourier Transform (FFT) with PBTG. Finally, combined with NIBC, it is seen that green function in the sea water is discretized with dense grid, e.g., 64 4

21 points per free space wavelength which the air green function is discretized with coarse grid, e.g., 8 points per free space wavelength. 1.5 Overview of the thesis The thesis focuses on numerical modeling of sea surface interaction with electromagnetic waves at L band. We compared the numerical results with other analytical method and studied an interesting phenomenon called Negative Upwind-Crosswind (NUC) Asymmetry for 3D problems. This is the first time that the numerical problems are used to compare to the measured data and used to study the NUC phenomenon [44]. For the fast algorithm, we have proposed Neighborhood Impedance Boundary Condition (NIBC) and combine SMCG and PBTG to accelerate computation speed and speed up the iteration rate. NIBC is proposed to solve the dense grid problem resulted by the large permittivity of sea water. The study of dense grid with more than 30 points per free wavelength is never reported in previous numerical models [47] [50]. The Aquarius data used in this thesis is provided by Dr. Simon Yueh from the Jet Propulsion Laboratory. 5

22 2.1 Introduction Sea Surface Radar Scattering at L-Band Based on Numerical Solution of Maxwell s Equations in Three Dimensions (NMM3D) Physical models of electromagnetic scattering from random rough surfaces have broad applications in microwave remote sensing of land surfaces and ocean surfaces. Analytical approximate methods include the small perturbation method (SPM), Geometrical Optics (GO), Physical Optics (PO), Advanced Integral Equation Method (AIEM), the Small Slope Approximation (SSA) and the composite surface model [1-6]. Ocean surfaces have a range of scales, with gravity waves of order of meters and short gravity and gravity-capillary waves with scales in the range of centimeters to millimeters. A common scattering model is the composite surface model, also known as the two scale model (TSM) [7-8]. In this approach, the surface height spectrum is divided in two in order to separate large- and small-scale roughness elements. Physical/Geometrical optics and SPM are applied, respectively, to the two scales, and the total radar cross section is obtained by summing the two contributions. The dividing line is usually chosen to be a half of the free-space wavenumber of the incident microwave [9, 10]. The two-scale model is a heuristic model as the decomposition of an object into parts followed by coherently or incoherently combining the scattering of each does not follow from Maxwell equations, otherwise Maxwell equations can be readily solved for any large object. Thus TSM requires tuning of dividing line between large scales and small scales. In recent years, numerical simulations of 3D rough surface scattering have become feasible by combining fast numerical methods and high performance computing [11-15]. The simulations 6

23 have been applied to land and ocean surfaces. Both land and sea surfaces have fine-scale (mm to m) features that can cause significant microwave radar backscattering. For example, gravitycapillary and capillary waves have small radii of curvature compared with microwave wavelength that can lead to strong backscattering and cross-polarized returns. Because of the fine scale surface features, the surface integral equation approach, with discretizations of surfaces, is more accurate than using the FDTD (Finite-Difference Time-Domain) or FEM (Finite Element Method) approaches. The latter two methods require volumetric discretization. The surface integral equation approach combines fine-scale surface discretization with the method of moment (MoM) to capture the impact of all relevant surface structure on the scattered electromagnetic radiation. Examples of fast dense matrix solvers include the sparse matrix canonical grid method (SMCG) [16] and the Multilevel Fast Multiple Algorithm (MLFMA) [12]. In order to reduce the number of iterations, a near-field preconditioner has been used [14]. Some past work in numerical simulation treated the ocean surface as a perfect electric conductor (PEC) [15]. Some have implemented impedance boundary conditions (IBC) [17, 41]. The impedance boundary condition states that the tangential electric field is equal to the product of the wave impedance of the ocean permittivity and the tangential magnetic field. The approximation is valid if the radius of curvature of the surface is large compared with the microwave wavelength. Compared to previous works which are making approximations in the modeling process, in this chapter Maxwell equations are solved with a finite permittivity. Some earlier works are simulating 1 dimensional surfaces or 2 dimensional surfaces with pulse basis functions, while in this work 2 dimensional surfaces with RWG basis functions are being studied [18-23]. In this chapter, we apply NMM3D (Numerical Maxwell Model in 3-D) [11] to simulate L- band radar scattering from a wind-roughened ocean. The surface integral equations of PMCHWT 7

24 (Poggio-Miller-Chang-Harrington-Wu-Tsai) [12] are used. The ocean surface profiles are stochastically generated using three dimensional ocean spectra. In PMCHWT formulation, two separate Green s functions are used, one for air permittivity and another for ocean permittivity. The Sparse Matrix Canonical Grid (SMCG) [16] is used in conjunction with RWG basis functions. For modeling of the rough surface scattering, the physical problem is 3D vector wave scattering. However, the rough surface boundary has 2 dimensions with 2 degrees of freedom. The SMCG method uses a Taylor expansion of the Green s function about the flat surface so that 2D Fast Fourier Transforms (FFT) can be applied to the 3D electromagnetic scattering problem. The SMCG method has been implemented in a parallel computation framework [24]. The scattering problems in this chapter include cases with up to 6 million surface unknowns in the tangential electric and magnetic fields. In NMM3D simulations, a finite surface size is used and the surfaces are discretized. This means that the ocean spectrum is truncated between k l and k u, where k l is inversely proportional to the surface size while k u is inversely proportional to the discretizing sampling. Note that in NMM3D simulations, there is no need to impose a surface scale separation within this domain of k l to k u. By changing k l or surface size, one can investigate the influence of adding roughness due to longer ocean surfaces. By varying k u, one can assess the impact of small scale capillary waves. This study will focus on L-band ocean backscattering simulations and their comparison to results obtained from the Aquarius satellite radar system for wind speeds of 5, 8 and 10 m/sec. Aquarius measurements provide estimates for VV, HH, VH, and the VV/HH ratio and NMM3D backscattering results are produced for each case. Comparisons will also be made with TSM backscatter estimates to evaluate the efficacy of this widely used approximation method. The chapter is organized as follows. In section 2.2, we describe the ocean spectrum used in this study. In section 2.3, we describe the formulations of surface integral equation and the 8

25 numerical implementations. In section 2.4, the influences of k l and k u, the lower and upper limits of truncated ocean spectrum on backscatter are illustrated. In Section 2.5 we illustrate numerical results and make comparisons to data and the TSM model. Also in this chapter, the Aquarius data is actually the results of geophysical model function (GMF). The GMF is the experimental formula that relates the backscatter with wind speeds and wind directions. The following is a typical definition of GMF, 1 1 cos 2 cos 2 pq A w A w A w (1) In equation (1), σ is the backscatter. w is wind speed. ϕ is wind direction. A, A 1, A 2 are coefficients through data fitting with measurement data. 2.2 Ocean spectrum and surface generation Ocean roughness spectra have been extensively used in electromagnetic modeling of ocean surface scattering. Common ocean spectra include the Pierson-Moskowitz, Durden-Vesecky [16, 26] (DV), Donelan-Banner-Jahne and Elfouhaily spectra [25, 27-29]. These wind-dependent ocean wave height models are written in the form of the product of an isotropic part and an azimuthal anisotropic (wave direction) part 1 W k,, u S k, u k,, u (2) k k,, u 1 k, ucos 2 (3) where S(k, u) is the omnidirectional spectrum with k being the spectral components and u being the surface wind speed. φ is the azimuthal angle with respect to the wind direction. Φ(k, φ, u) is the directional spreading function or angular dependence function which denotes the nonuniform distribution of wave roughness in azimuth. In this chapter, we used Durden-Vesecky spectrum for 9

26 the NMM3D simulations. Appendix B gives the expression of the Durden Vesecky spectrum. Two versions are used. One version is the isotropic, or omnidirectional, spectrum defined by setting Δ(k, u) = 0, and is referred to as isotropic DV in this chapter. The other version used is the full directional DV spectrum, referred to as the anisotropic DV spectrum. Although other spectrum can be used such as the Elfouhaily spectrum, the purpose of this chapter is to use NMM3D simulations to compare with the results of TSM and small perturbation method. Thus we limit ourselves to the isotropic and anisotropic DV spectrum [26]. In this study, the microwave frequency is 1.26GHz, which is the frequency of the Aquarius L-band scatterometer. The wavenumber of the microwave k 0 is 26 m -1. According to the Bragg scattering theory [29, 30], the spectral component of the ocean spectrum that contributes to the backward scattering is k 2k sin B (4) 0 where for the incidence angle ( ) of 39, the Bragg scattering wavenumber is 33 m 1. In the rough surface simulations computed using DV as input for the surface integral equation approach [15], we specify the surface area to be of an extent L x and L y in x and y directions respectively. We also set surface lengths L x = L y = L. In the discretization of the surfaces, sampling intervals of Δx = Δy are used. These limits set surfaces that contain spectral components W(k, φ, u) in the range of k l k k u, where k u (5) x k l 2 (6) L 10

27 In Figure 1, we show a log-log scale of the isotropic Durden-Vesecky spectrum for wind speeds 5 m/s, 8 m/s, and 10 m/s. The wavenumber of the incident microwave, the Bragg wavenumber, k l, k u for several surface lengths and sampling are also shown. In Figure 1, the lower limit k l corresponds to a surface by 64λ 64λ and the upper limit k u corresponds to x = y = 1/32λ. The incident L-band and Bragg wavenumbers lie between the upper limit and lower spectral limits. In Table I, we tabulate k l and k u values corresponding to the surface sizes and the discretization intervals used in this chapter. With the increase of surface length, large scale ocean waves are incrementally increased in the simulations. The peak wavenumber for 5 m/s, 8 m/s and 10 m/s are around 0.27m 1, 0.1m 1, and 0.07m 1. With a limited surface simulated, not all scales of ocean waves are included in the simulation. The rms heights for wind speed 8m/s 10 m/s are not the same as real ocean surfaces. Figure 1 Isotropic Durden-Vesecky surface elevation spectrum (solid lines) for wind speeds 5 m/s, 8 m/s, and 10 m/s. Magenta and blue lines denote lower and caupper limit. The lower limit k l corresponds to a surface size of Upper limit k u corresponds to x y 1/ 32. Red 11

28 and green lines denote the incident microwave wavenumber and Bragg wavenumber respectively. Table I Upper and lower spectral limits for varying sampling point densities and surface sizes Surface x, y (λ) k l (m 1 ) k u (m 1 ) Length (λ) 16 1/ / / / PMCHWT surface integral equation and numerical implementation using RWG and SMCG In this section, we summarize the methodology of NMM3D simulations [13-15]. For scattering by dielectric surfaces, the PMCHWT formulations of surface integral equations (SIE) are given in the following form: 1 T T1 K K1 inc J n E inc M n H K K1 T T 1 1 (7) where T X ik ˆ 2, ' ' ' i in I g S i r r X r dr ki (8) X X K nˆ PV.. g r, r ' r ' dr ' i S i (9) 12

29 and g i r, r ' iki r r ' e 4 r r ' (10) 1 iki r r ' exp iki r r ' gi r, r ' r r ' 4 r r ' (11) k In the equations above, and g r, r ' are the wavenumber and Green s function in air, respectively, while and g1 r, r ' are the wavenumber and Green s function for ocean k 1 permittivity. Next we apply the MoM to the above surface integral equations. RWG basis functions are used. Curl-conforming RWG functions are used as testing functions. A property of the discretized equations is that the impedance matrix elements are symmetric, e.g. Z mn Z nm (12) We have used this property to reduce CPU memory and computation time. A finite surface area of L by L is used in the surface integral equations. To eliminate the edge effects for a surface with finite size, a tapered plane wave [29], in the form of 2D Fourier transforms, is used: kzz inc i k E r dke k e k (13) k zz k inc i k H r dke h k (14) where r zz ˆ (15) 13

30 k xk ˆ ˆ x yk (16) y and 2 2 k k, 0 k k z z 2 2 i k k, k k k k k ˆ ˆ ˆ ˆ i e k e h k h k v k v k ˆ ˆ ˆ ˆ i h k e v k h k h k v k (17) (18) (19) where ˆ ˆ e e k E h k E v k (20) i i h i v i The notations xˆsin ˆ i y cos i, k 0 ˆ hk 1 xk ˆ y yk ˆ x, k 0 k v k kz kk xˆcos yˆsin, k 0 i i k xk ˆ ˆ x yk y zˆ, k 0 k (21) (22) For fast matrix vector multiplication in the iterative solution of the matrix equation, we used the SMCG method. In the SMCG a Taylor series expansion of Green s function is applied [16] about a flat surface defined by the x-y plane. Subsequently a 2D FFT is applied. The Taylor series expansions are performed for the Green s functions and the gradient of the Green s functions, 14

31 iki r r ' M 2 e 1,2, ' z d gi r r am R 2 4 r r ' m0 R 2 (23) 1 ' exp ' M 1,2 2 iki r r iki r r z d gi r, r ' r r ' bm R 2 4 r r ' m0 R 2 (24) 1,2 1,2 where z f x, y f x', y '. The coefficients above, a and b, are d translationally invariant in the horizontal directions. We apply 2D FFTs to the coefficients and these are performed for the product of the Toeplitz matrix and column vector [16]. The first four coefficients of the above expansions, a and b, are listed in Appendix A. In Appendix A, we m m also describe improvements to use of SMCG on NMM3D compared with that used in previous implementations. In this study, the Taylor series are expanded to fifth order. We have checked that 5 terms in Taylor series expansion could give accurate results compared to direct green s function multiplication such that the error norm is less than 0.2% for wind speed 5 m/s. The details of numerical implementation of SMCG on RWG basis functions are also described in appendix A. m R m R 2.4 Ocean surface profile generations and L-band backscattering variations with k l and k u For L-band NMM3D simulations, we account for roughness due to both short gravity waves (5 m and less) [32] and shorter gravity-capillary waves (down to 2 cm or so) [33]. Since the wavelength of wind-generated gravity waves will increase with wind forcing, a larger sea surface for the NMM3D analysis is necessary for greater wind speeds. To account for the influence of fine capillary waves on backscatter, a small discretization of the surface is necessary. By varying surface length and discretizations as shown in Table I, we examine the effects of longer gravity 15

32 waves and shorter gravity-capillary waves on backscattering at L band. At wind speed 5 m/s, Aquarius data show the Negative Upwind-Crosswind (NUC) asymmetry [33, 37]. This means upwind backscatter is smaller than crosswind backscatter. At other wind speeds (>=8m/s), upwind backscatter are stronger than crosswind. This asymmetry cannot be reproduced based on the DV anisotropic spectrum or using other existing spectrum. We will investigate the spreading function in the future. In the meantime, we use the isotropic DV spectrum for NMM3D for 5 m/s, but upwind data for Aquarius. For 8m/s, the Aquarius data is weakly dependent on wind direction, especially for beam 1 and beam 2. Then we used isotropic surfaces for 8m/s. For 10m/s, we used anisotropic surfaces. We have also calculated surface sample statistics from randomly generated surface profiles [34-36]. The rms heights are provided in Table II for various surface lengths and wind speeds. We have made comparisons between numerical computation and analytical integration and found the ocean statistics within the same spectrum range are the same between the two methods. Table II Rms for isotropic DV spectrum at different wind speed in terms of longest surface wavelength Ocean Surface k l (m 1 ) rms wind speed (m/s) Length L(λ) in centimeters

33 For wind speed of 8 m/sec, the increase of surface length from 16 wavelength to 64 wavelength, corresponding to the decrease of k l, results in an increase of rms height from 5.1 cm to 13.8 cm. This indicates inclusion of longer surface waves with increased surface size. In order to study the influence of the upper and lower limit of the spectral range used in the simulation, we conducted NMM3D simulations using different surface lengths and discretizations. Table III VV and HH for varying discretizing densities at wind = 5 m/s with 39 incidence angle and isotropic DV spectrum compared with upwind Aquarius data x, y (λ) k u (m 1 ) VV(dB) HH(dB) # of realizations 1/ / Aquarius N/A N/A Table III shows the backscatter results for wind speed = 5 m/s at several sampling densities. The results show that 16 points per wavelength and 32 points per wavelength yield very similar backscattering results. The difference is within 0.3 db. The Aquarius upwind data are also listed in the table. Based on extensive simulations, sampling at 16 points per wavelength yields sufficient accuracy to account for the small scale roughness. Including smaller capillary wave roughness does not alter the backscattering results for this case. For emissivity calculations, denser sampling 17

34 are required. And results for cross wind and upwind are compared for the anisotropic DV spectrum and are shown in Figure 5. Next we conducted numerical simulations for several surfaces sizes at winds of 5 and 8 m/s. These correspond to changing k l. The backscatter results are shown in Table III. We also include in Table III and IV the number of realizations used in the simulations as shown in the last column of the two tables. We have conducted convergence test with the number of realizations. Results show that 30 realizations can give stable results for backscatter. For the 16λ case, the computation time is about 3~4 minutes for each realization. For the 32 λ case, it is about 20 minutes for each realization. For the 64λ case, it is about 1 hour for each realization. The convergence rate of iterations is influenced by the roughness of the ocean surface. Table IV VV and HH for L band 39 incidence isotropic DV spectrum compared with Upwind Aquarius data (a) VV and HH for varying surface sizes at wind=5 m/s Surface Size k l (m 1 ) VV(dB) HH(dB) # of realizations 16 16λ λ λ Aquarius N/A N/A (b) VV and HH for varying surface sizes at wind= 8 m/s Surface Size k l (m 1 ) VV(dB) HH(dB) # of realizations 18

35 16 16λ λ λ Aquarius N/A N/A The NMM3D backscatters of various simulated surface sizes are shown in Table IV (a) and (b) for wind 5 m/s and 8 m/s, respectively. It is noted that there is a less than 1dB difference for VV, a 1.4 db difference for HH at 5 m/s and 8 m/s. One observes that for the two wind speeds, the maximum difference for backscatter for varying surface sizes is 0.7 db. There is also an increase of backscatter with increasing wind speed from table (a) to table (b). It is to be noted that from table (c), the 3 surface sizes at wind 8 m/s have approximately the same backscatter. But in Table II, the 3 surface sizes have very different rms height for the same wind speed of 8m/sec. With the increase of simulated surface size, longer waves (waves that are meters long and have a large rms height) are included in the simulations. However, these longer waves do not cause significant differences on backscatter at 8 m/s. The backscatter arises from the spectral components in the ocean spectrum that are comparable to the wavelength of the incoming microwaves. For L band, the decimeter and centimeters ocean waves cause backscatter. The above convergence analysis may also be influenced by the random errors brought by the Monte Carlo simulations. In future studies, the use of larger surface sizes will be studied. 2.5 NMM3D results and discussions In this section, we present NMM3D numerical ocean scattering results at L-band and their comparison with Aquarius data. The assumed surface relative permittivity is 75+61i, corresponding to a complex index of refraction of i. This is the case of seawater with a salinity of 35 ppt (parts per thousand) at water temperature of 10 o C at 1.26 GHz. For each 19

36 realization, a single sea surface is generated and Maxwell s equations are solved numerically. The scattering coefficients represent results averaged over 30~300 realizations. Isotropic and anisotropic Durden-Vesecky ocean spectrum are used for simulations of wind-dependent L-band backscatter. The isotropic DV spectrum is used for 5 m/sec and 8 m/sec. The anisotropic DV spectrum is used for 10m/sec [38, 40]. Figure 2 Comparison of L-band NMM3D scattering amplitude distributions for 300 realizations for co-polarized cases VV (left), HH(right) and for wind speed of 5 m/s using the isotropic DV spectrum and an incidence angle 29 o. The red curve indicates that for a Rayleigh-distributed process. Figure 2 shows the statistical distributions of the simulated results for HH, VV computed using 300 realizations with a wind of 5 m/s and incidence angle of 29 o. The histogram of computed backscatter values is consistent with that shown for a Rayleigh distribution. Figure 3 shows the in-plane bistatic L-band scattering coefficients at 5 m/s wind speed and 39 degree incidence. In these simulations we decompose the bistatic scattering coefficients into coherent and incoherent waves [13, 15]. For the surface sizes used, the coherent waves are only significant in the specular direction while scattering in other directions arises from the incoherent waves [13-15, 39, 40]. The backscattering direction is highlighted with a dashed line in Figure 3. 20

37 Figure 4 shows the backscatter dependence at the three incidence angles of the Aquarius scatterometer, 29, 39, 46 at wind = 5 m/s. Backscattering decreases with incidence angle. The upwind Aquarius data are also shown. For co-polarization, the agreement between simulations and upwind Aquarius data are within 1 db at the three incidence angles. Figure 3 In-plane bistatic VV HH for wind = 5m/s isotropic DV spectrum Figure 4 L-band backscatter at wind speed = 5 m/s for different incidence angles: 29, 39, 46 degrees and isotropic DV spectrum compared with upwind Aquarius data. 21

38 In Figure 5, we illustrate the backscatter for wind = 10 m/s using an anisotropic DV spectrum. The incidence angle is 46. There are 4 curves in the figures: NMM3D with error bars derived from the Monte Carlo simulations, first order SPM, Aquarius data, and composite surface model (two scale model). There are several TSM models because TSM is a heuristic approximation, subdividing the objects into parts and them performing incoherent combining with tilts. We have referred to several TSM models in our computation. The uncertainty of numerical simulation is defined as 1 error ( db) 10log10(1 ) (25) N for VV, HH. In (25), N is the number of realizations. The 4 panels in Figure 4 correspond to VV, HH, VV/HH (polarization ratio), VH, respectively. We can see that HH of SPM is lower than Aquarius data, while NMM3D, TSM, and SPM show VV results that are all comparable with Aquarius observations. The polarization ratio, VV/HH, derived using SPM is 1.5dB larger than the data. NMM3D and TSM can both give good polarization ratio compared to data while polarization ratio given by SPM is larger than data and not changing with data. For current wind speeds (wind speed<10m/s) we are studying, TSM and NMM3D both agree well with data for cross polarization. The figure also indicates that Aquarius data show a directional dependence in the polarization ratio. A wind direction dependence for this ratio is also exhibited in NMM3D, while the ratio is direction-invariant for SPM. With a diving wavenumber equals k 0, we see that 2 TSM can predict good scattering coefficient as NMM3D. Figure 6 provides the backscattering polarization ratio for winds of 8 and 10m/s and for the three incidence Aquarius angles [40]. NMM3D simulation uncertainties are also shown. For 22

39 wind=8m/s, the surfaces used in the computations are derived from the isotropic DV spectrum. The results of TSM and data shown in Figure 6 are averaged over all wind directions. For wind=10m/s, the anisotropic DV spectrum is used in NMM3D. The results of NMM3D, TSM and Aquarius data are also averaged over all wind directions. Results indicate that the VV/HH ratio increases with increasing incidence angle. From the figures, we see that the difference for both TSM and NMM3D compared to Aquarius data are both within 0.6dB. Figure 5 L-band backscatter for wind speed =10 m/s and incidence angle= 46. Upper two panels are VV and HH, and bottom is for VV/HH. TSM, NMM3D, SPM results use the anisotropic DV spectrum. 23

40 Figure 6 Comparison of NMM3D and two scale model of VV/HH with Aquarius at 29 o, 39 o, and 46 o. NMM3D is using isotropic DV spectrum for 8 m/s and anisotropic DV spectrum for 10 m/s. TSM is derived using the anisotropic DV spectrum for wind = 8 and 10m/s. Each data point is obtained by averaging over all wind directions. Figure 7 Polarization ratio for wind speed 10m/s and 26 incidence. The horizontal axis is different cutoff wavenumbers. Vertical axis is VV/HH ratio in db. Each data point is obtained by averaging over all wind directions. 24

41 In Figure 7 we plotted the polarization ratio for wind speed 10m/s and 26 incidence. The horizontal axis is different cutoff wavenumbers. Vertical axis is VV/HH ratio in db. Each data point is obtained by averaging over all wind directions. It is apparent that the polarization ratio is changing with different cutoff wavenumbers. And for such a case, we see that k d = k 0 will result in a VV/HH ratio 3dB higher than k d = k Conclusions With an efficient SMCG method implemented to solve the PMCHWT surface integral equations, we have performed NMM3D simulations for ocean radar scattering at L band frequencies. Surface areas used are up to 15.2 m by 15.2 m (64 wavelengths). For the largest surfaces, there are more than 6 million surface unknowns in the tangential electric and magnetic fields. This use of a large surface for simulations allows us to examine the effects of gravity and capillary waves. The numerical simulations, unlike the two-scale scattering model, does not require a wavenumber dividing line between large and small scale roughness. Simulations are shown for 5m/sec, 8m/sec and 10m/sec using isotropic DV spectrum and anisotropic DV spectrum. The backscatter results are in good agreement with Aquarius observation data for VV and HH copolarization, cross-polarization, and VV/HH ratio. A rigorous approach to the problem based on NMM3D is to incrementally increase the surface size until convergence is reached. Recent computational electromagnetic efforts about scattering by large objects are on using rigorous domain decomposition which should be considered in remote sensing simulations. Without a large domain, it is possible that tilt effects are not included. To obtain tilt effects require a definition of decomposition and tilts can both decrease or increase local incident angles, requiring a heuristic use of tuning parameters. For inclusion of tilt effects, the readers can refer to reference [7]. For NMM3D to move forward, the increase of domain size via domain decomposition and the increase 25

42 of sampling points per wavelength are essential steps. Also, the authors has used k 0 2 as the dividing wavenumber in the TSM for separation of applying SPM and Geometrical Optics. The results produced by TSM match well with Aquarius. But if using other dividing wavenumbers, the scattering coefficients are not good. This indicate TSM is quite an ad hoc method. 26

43 Appendix A In order to integrate SMCG with the PMCHWT equation, it is necessary to write the equation in a form where the field and source points are separated. We will use the subscript m to denote the field point and subscript n to denote the source point. The superscripts x, y, and z represent the three axes. The following discretized impedance matrix expressions are obtained after performing moment method operations to the kernel of the electric field integral equation (EFIE) Sm Sn 1 fm fn f ' 2 m fn g r, r ' drdr ' k (26) where f and f are the RWG basis functions for the field and source edges. Also m n x x, ' ', ' ' y y fm r g r, r ' fn r ' z z f r g r, r ' f r ' f r g r r f r f r g r r f r m n m n m n (27) l l, ' ' m n fmg r r fn g r, r ' Sm Sn (28) The kernel of the magnetic field integral equation (MFIE) is: m, ' ' f r g r r f r n 1 ik r r ' exp ik r r ' ' ' fm r r r fn r 4 r r ' (29) We use G r, r ' 1 ik r r ' exp ik r r ' 4 r r ' (30) Then 27

44 fmr g r, r ' fnr ' fmr r r ' fnr ' G r, r ', ', ', ', ', ', ', ', ' x z y y z x x y z f r G r, r ' f f r G r, r ' f f G r, r ' f r f G r, r ' f r G r r f f r G r r f f G r r f r f G r r f r y x z z x y y z x z y x m m n m m n m n n m n n f r G r r f f r G r r f f G r r f r f G r r f r z y x x y z z x y x z y m m n m m n m n n m n n m m n m m n m n n f r y x z m n n (31) Equation (27) (28) and (31) are in a form such that the source point is on the right side of Green s function while the field point resides on the left side of Green s function. Next the Green s function is expanded in a Taylor series with respect to the flat surface. The first 4 coefficients are listed in equations (32) to (39). a 1,2 0 R 1ik1,2 R exp ik 1,2 4 R R (32) a 2 1,2R exp ik k 3ik 3 1,2 1,2 1,2 1 R R 2R 2R 2R (33) a 3 2 1,2R exp ik ik 6k 15ik 15 1,2 1,2 1,2 1,2 2 R R 8 8R 8R 8R (34) a ,2R exp ik k 10ik 105k 35ik 35 1,2 1,2 R 1,2 1,2 1,2 3 R R R 16R 16R b 1,2 0 R exp ik1,2 R 4 R (35) (36) 1 R expik1,2 R b 1,2 1,2 ik R (37) b k 3ik R 2 1,2 1,2 R 1,2 2 R expik1,2 R (38) 28

45 ,2 k1,2 R k1,2 R 13ik1,2 13 b3 R expik1,2 Ri R (39) We have used the centroid as the integration point within each triangular patch [26]. A C B D (a) (b) Figure 8 SMCG with centroid as integration points As is shown in Figure 8 (a), the local coordinates of the two integration points A and B are (1/4, 3/4) and (3/4, 1/4) which are not centroids of the two triangles. In Figure 8 (b), the local coordinates of C and D are (1/3, 2/3) and (2/3, 1/3), respectively. The points C and D fall onto the center of each triangle. After making the modification from (a) to (b), the accuracy of SMCG is greatly improved. We define the error of SMCG by error z mn z z mn SMCG mn (40) where z mn are the impedance matrix elements calculated directly using 7 th order quadrature SMCG and z are the matrix elements calculated indirectly by the SMCG method. For a 4 by 4 mn 29

46 surface with 16 points per wavelength, the error is about 0.2%. In addition to SMCG, we have also used Green s function interpolation in the numerical simulations [9]. g r, r m n r m r n Figure 9 SMCG in conjunction with RWG basis functions In Figure 9, we illustrate how SMCG based on RWG functions works. Firstly, the currents inside one source triangle patch aggregates to the centroid. Next, the accumulated currents transfer to the field triangle through Green s function. Then the transferred field distributes to each inner edge inside the source triangle. 30

47 Appendix B The ocean surface height spectrum for a fully-developed wind sea, as proposed by Durden and Vesecky [21], has the following form: 2 g exp 2 4 a ku WDV k. 4 DV k, alog 10 k/2 2 k 2 bku* g* k k k k j j (41) 1 where k 2m. The subscript DV denotes the Durden-Vesecky spectrum. The constant j a0 is in [21], but was adjusted upwards to in Yueh et al. [20]. In order to best fit the L-band measurement data in this study, a value of a is applied. Other parameters are , b 1.25, a 0.225, g g k, g 9.81 m/ sec is acceleration due to gravity, and * m / sec is the ratio of surface tension to water density. The wind friction velocity at the ocean surface, u *, is related to the wind speed at a height (h) of 19.5 meter above sea level, U 19.5, using the following form. U h u* h log / u u * * (42) The wave directional spreading factor k, is (43) 2 ( k, ) 1 ( k)(1 exp( sk ))cos(2( wind )) o In (43), the upwind direction is defined as 0. In this chapter, k wind is defined as k 1 R 2 c 1R 1D (44) with 31

48 R U (45) x10 U12.5 D 0 dkk S k expsk dkk S k 0 (46) 32

49 3.1 Introduction Radar Scattering of Ocean Surfaces with Anisotropic Ocean Spectrum using NMM3D simulations Recent years have seen an increase in L-band microwave satellite sensors in earth remote sensing applications. This wavelength benefits from near atmospheric transparency, the ability to retrieve soil moisture information, sense and penetrate snowpack, and respond to ocean surface wind and salinity [45]. Studies have recently been conducted for ocean backscattering at L band using PALSAR and Aquarius radar data [38]. Because ocean wind create anisotropy in ocean surfaces, the radar cross section have interesting azimuthal directional dependences. There is a new puzzle for L band in directional dependence in all polarizations: the upwind backscatter is larger than the crosswind backscatter for high winds (>10 m/s), but the relative amplitude changes at lower wind speeds (<8 m/s). This is termed a Negative Upwind-Crosswind (NUC) asymmetry. Figure 10 shows the Geophysical Model Function (GMF) of Aquarius scatterometer for different wind direction. The incidence angle is 29 degree. The figure shows that for wind speeds larger than 8 m/s, backscatter of crosswind is smaller than that of upwind. But for wind speeds of 5 m/s and 8 m/s, the backscatter at crosswind is larger than that of upwind. This phenomenon is also shown in PALSAR [35] which is L band. However it not observed in higher frequency bands. Analytical models, AIEM or TSM, using common adopted ocean spectrum, cannot reproduce the NUC phenomenon. Yueh et al. (2013) suggested that directional and passive/active aspects of the signals might correspond to a non-bragg scattering phenomena at L-band. 33

50 Alternatively, it is possible that the surface waves themselves have markedly different directional distribution near the L-band, ultragravity cm wavelengths, at these low to moderate wind speeds. 34

51 Figure 10 Geophysical Model Function of Aquarius scatterometer data versus wind direction of beam 1 at L band. Top figure is VV, middle figure is HH, bottom figure is VH. The four curves correspond to wind speeds 5, 8, 10, 15 m/s. In addition to analytical modeling of surface scattering problems, numerical simulations of 3D rough surface scattering have become feasible by combining fast numerical methods and high performance computing. The simulations have been applied to land and ocean surfaces[11, 12, 44, 46]. In [44], it is proved that the NMM3D results could predict the backscatter well compared to Aquarius data and analytical method. On the other, the input for the numerical method is the computer generated ocean surfaces which can include a large range of different ocean wave scales. This is advantageous compared to many analytical methods which only one point on the ocean spectrum is used. In this chapter, we study the NUC through both analytical method and numerical simulations at L band under 5 m/s. Two kinds of surfaces are studied: Elfouhaily spectrum with/without horseshoe patterns and the Apel spectrum with modified angular spreading function [43]. In the following ocean spectrum is introduced and then results generated from NMM3D are shown and analysis are conducted. 35

52 3.2 Ocean spectrum Ocean spectrum is used to characterize the power distribution of ocean waves on different scales of waves and on different directions. In the past we studied soil surfaces with exponential correlation functions [11, 46].Compared to rough land surface, ocean surface is smoother. This is also illustrated in Figure 11. Figure 11 shows a comparison of spectrum between ocean (Pierson-Moskowitz), spectrum of Gaussian and exponential correlation function. The three curves have the same rms height which is 0.06 meter. The correlation length is 10 times rms height. It is clear that exponential function has a much higher tail than ocean spectrum. This means soil land surfaces have much finer features than sea surface. Gaussian surface is even smoother. Figure 11 Comparison of Ocean (Pierson-Moskowitz), Gaussian and Exponential Spectrum for the same RMS height. Correlation length for Gaussian and exponential spectrum density is 10 times rms height. 36

53 From ocean spectrum, the sea surface profiles are generated using spectrum for varying wind speeds using linear and nonlinear methods. The basic assumption underlying linearity of the surface is that the surface is a linear combination of different scales of sinusoidal waves. Figure 12 plots the computer generated ocean surface using linear method by Elfouhaily spectrum. The dimension is 15.2m by 15.2m. The wind speed is 5m/s. In this chapter we simulate L-band backscattering from anisotropic ocean surface with NMM3D. Two kinds of ocean surface profiles are used for NMM3D simulations. One is a group of time series nonlinear profiles generated by combining the nonlinear gravity waves and shorter horse-shoe pattern using Elfouhaily spectrum. The other is a group of linear profiles generated from a newly developed directional spectrum [43]. The introductions and analyses of these two kinds of profiles are given in the following sections. Figure 12 Computer generated linear ocean surface at 5 m/s using Elfouhaily spectrum. The dimension is 15.2m by 15.2m. 37

54 3.2.1 Nonlinear ocean surface profiles In addition to linear ocean surfaces, non-linear ocean surface is more popular due to its matching with ocean hydrodynamics. The profiles we are using is the Choppy Wave Model (CWM) with/without horsehose patterns. Figure 13 Extracted ocean spectrum for 5m/s with and without horseshose patterns of Choppy Wave Model for upwind and downwind. Figure 13 shows the extracted ocean spectrum for 5m/s with and without horseshoe patterns of Choppy Wave Model. The ocean spectrum used is anisotropic Elfouhaily spectrum. The profiles are of time series with 0.1 second as the time interval of total 2 seconds. It is seen that for k around the Bragg wavenumber, upwind is stronger than crosswind. Generally the upwind is stronger than crosswind for all the wavenumber ranges. The surface length of the profiles are 8 meters A newly developed directional wave spectrum Another surface profiles that have been used is linear ocean surface generated by Apel spectrum using the spreading function by Du et al., 2017 [43]. The angular spreading function is obtained by data fitting between Aquarius measurement and AIEM using Apel spectrum. Figure 14 illustrates the (k) function of Du s spectrum for the anisotropic surface for 5 8 and 10 m/s. We could see that different from many other spectrum, the (k) function is negative for k ranges from 10 to around 80 rad/m. If using first order SPM to simulate backscatter, we will 38

55 get the backscatter is larger in crosswind than upwind at L band. This is because the input of first order SPM is the Bragg point in the spectrum curve. And we have Bragg wavenumber equal 33 rad/m for 39 incidence and 1.26GHz. Then 1 33cos(0) 1 33cos(2 ) 2 (47) The new angular spreading function (ASF) is expressed as k ( k, ) 1 cos 2 (48) 0 k tanh a a S a c / c a c / c d d p p m m (49) g 20 S tanh u k g u 2.55u k k c * *2.55 * d x 2 x m cp kx cp (50) * u u u10 (51) a m * u 0.26 (52) c m 39

56 Figure 14 (k) of Du s spectrum for wind speed 5 8 and 10 m/s. In equation (49), c is the wave phase speed and c p is the phase speed of the dominant long wave. The relationship between wave phase speed c and wavenumber k can be written as 2 k c g / k 1 k m (53) a ln 2 / 4 a ln 2 / 2 a c 0.23 m / s (54) 0 d p4 m 40

57 Figure 15 Comparison of upwind and crosswind spectrum with the new spreading function by Du and Apel spectrum for 5m/s. Left figure is the zooming in of the right figure for k corresponds to 10 rad/m to 50 rad/m. 41

58 Figure 15 shows the Comparison of upwind and crosswind spectrum with the new spreading function by Du and Apel spectrum for 5m/s. It is observed that upwind spectrum is higher than crosswind for whole wavenumber range except for millimeter waves and centimeter waves which corresponds to k equal several tens of rad/m. Even though two kinds of spectrum are used for comparison: Apel and Elfouhaily. The purpose of this chapter is to explore the reason of NUC, and NUC is a comparative relationship between upwind/downwind backscatter and crosswind backscatter. So even though different spectrum will create different absolute VV HH value, we pay more attention on the comparison between upwind and crosswind and also VV/HH ratio. 3.3 Results and discussions We use two kinds of ocean spectra in this section: Elfouhaily spectrum with or without horseshoe features and Apel spectrum with modified angular spreading function. Three results are compared: experimental data from Aquarius, analytic theoretical results of AIEM and 3D numerical simulations of NMM3D. Figure 16 shows the in-plane bistatic L-band scattering coefficients at 5 m/s wind speed and 46 degree incidence using Apel spectrum and modified angular spreading function. In these simulations we decompose the bistatic scattering coefficients into coherent and incoherent waves. For the surface sizes used, the coherent waves are only significant in the specular direction while scattering in other directions arises from the incoherent waves. The backscattering direction is highlighted with a dashed line in Figure

59 Figure 16 Incoherent bistatic scattering coefficients by NMM3D for 5 m/s at L band using Apel spectrum and modified angular spreading function. Incidence angle is 38 degree Scattering of nonlinear Elfouhaily surfaces Table V VV and HH of surface w/o horseshoe using NMM3D (a) upwind Without horse shoe With horse shoe VV dB dB HH dB dB (b) crosswind Without horse shoe With horse shoe VV dB dB HH dB dB The surfaces used in this part are 20 surfaces in a time series: 2 seconds with 0.1 second as the time interval. So in fact there is a close relation between the different surfaces. From the results in Table V we see with or without horseshoe feature, the simulation results do not reproduce NUC. 43

60 And also the absolute values of the scattering differ significantly from the Aquarius measurements. This is mostly due to the fact that the 20 surfaces used are closely related to each other Scattering of Apel spectra surface with modified angular spreading function Figure 17 Comparison of backscatters of AIEM, NMM3D, Aquarius at L band for VV, HH and VH for different wind directions. First row: 29 incidence; Second row: 39 incidence; Third row: 46 incidence. The columns are VV HH and VH returns, respectively. Figure 17 illustrates the results by NMM3D, AIEM, and Aquarius measurement data. The spectrum used in this part is Apel spectrum with the modified ASF described in previous sections. From the figures, we see that the AIEM compared well with Aquarius. This is an obvious conclusion since the new ASF is obtained by data fitting between AIEM and Aquarius. For beam 2 and 3, NMM3D also displays the NUC feature. For beam 1, the like pol of upwind and crosswind are identical to each other. Overall, VV and HH of NMM3D predictions are within 2dB difference 44

61 of Aquarius measurement. And the crosspol prediction of NMM3D also has a maximum 2.4dB deviation from Aquarius. Figure 18 Polarization ratio comparison between NMM3D, AIEM, Aquarius for upwind for the 3 beams: 29, 39, 46 at wind speed 5m/s. Figure 18 plots the polarization ratio of NMM3D, AIEM, Aquarius for the 3 beams of Aquarius. As we can see from Figure 18, the maximum difference for the 29, 39, and 46 at wind speed 5m /s is 1dB between NMM3D and Aquarius. Through Figure 17 and Figure 18, we can see that NMM3D predictions are accurate, particularly for cross polarization. 3.4 Conclusions Recent advances in computers have enabled the simulations of sea surface scattering using numerical method through parallel computing in NMM3D. By implementing moment method with RWG basis functions in surface integral equation, simulation are conducted with the computer generated surface profiles as input. Through this, we are able to test the influence of different scales of ocean waves and different wave patterns on scattering. Also in this chapter an effort is made to explore the reason for NUC appeared at L band. Through numerical simulations, it is seen that horseshoe patterns cannot generate the NUC 45

62 asymmetry at 5m/s. We are proposing a possible reason leading to this: the centimeter waves and millimeter waves are stronger in crosswind than upwind. Through NMM3D simulations, we have seen that Apel spectrum with such a new angular spreading function can lead to backscatter in crosswind exceeding upwind. Thus it is maybe due to the stronger millimeter and centimeter waves at crosswind than upwind that lead to NUC. This is in contrast to one s intuition since wind is blowing in the upwind direction and naturally ocean waves should be stronger in upwind than crosswind. But recent spectrum measurement also shed some light on this reversal [36]. The ocean dynamics is now still being a hot topic due to its complexity. Up to now, there are various studies about ocean spectra. And the actual process of interaction between electromagnetic waves and ocean surfaces is also quite complicated. We will continue the study of numerical method on wave scattering due to ocean surface. 46

63 Scattering of Lossy Dielectric Surfaces in Full Wave Simulation of Maxwell s Equations with Dense Grid and Neighborhood Impedance Boundary Conditions 4.1 Introduction Lossy dielectric surfaces are characterized by moderate relative permittivity of about 5 to 80. The imaginary part is smaller than the real part of about 20% to 50% of the real part. Scattering by rough surfaces of these dielectric surfaces are often approximated by perfect electric conductor (PEC) or impedance boundary conditions (IBC). The exact formulation, on the other hand, should be based on dual integral equation with two Green s function g 0 which is the free space Green s function with the wavenumber k of the medium above and g 1 which is the Green s function with the wavenumber k 1 of the lossy dielectric. Compared to IBC or PEC, the dual surface integral equations (SIE) usually have more number of unknowns [47] due to that the unknowns are both electric field and magnetic field on the surfaces. But in order to achieve an accurate results, it is preferable to solve dual surface integral equations rather than solving PEC or IBC which are basically approximations. For ocean surfaces, as illustrated in Chapter 3, the ocean spectrum is smoother than soil surface. However, compared to soil, the sea water has a much larger permittivity. In previous chapters, the problem under consideration is sea water of permittivity with 75+61i with corresponding complex refractive index i. Soil often has a permittivity ranges from 1 to 30. Such a large permittivity of sea water will require a very dense grid to get correct results. This is because the surface electric currents are fluctuating frequently on the surface. The wavelength inside the sea water is around 8 times shorter than that in the air. For PEC scattering problem, 10 47

64 points per free space wavelength are usually used in MoM. For lossy dielectric case, if 10 points are used in MoM, this corresponds to we have 1.5 points per wavelength in dielectric medium. Apparently this is not correct. Thus to obtain correct results, around 60~100 points per free space wavelength in discretization of the surface is a must to get correct currents distribution. Table VI A comparison between PEC, IBC and dual SIE L=100λ PEC IBC Dual SIE dielectric No Yes Yes g 0 Yes Yes Yes g 1 No No Yes Δx N=number of points Number of unknowns λ 10 λ 10 λ In Table VI we are showing a comparison between PEC, IBC and dual SIE for a two dimensional problem. The case is 1 dimensional surface with length L=100λ. For PEC, the dielectric medium is basically considered to be a perfect electric conductor and as is mentioned 10 points per free space wavelength is enough. For IBC, the surface electric field and magnetic field are correlated through a simple relationship and only green function of the air is used. Thus 10 points per wavelength is also enough. For dual SIE, if 64 pints per wavelength is used we will have number of unknowns. Dual SIE is the only method among the three which has used the green function in the lower medium. In contrast, in implementation of MoM, in addition to much CPU and memory burden, a tricky problem of slow convergence arises when dense grid is used. With limited computational 48

65 resources, it is much preferable to come up with methods to reduce the iterative steps. In computational electromagnetics, 3 methods are mainly used to solve this problem. 1) change basis function 2) use preconditioner 3) use a different equation In this chapter, we are proposing a novel method to alleviate the problem of slow convergence with dense grid used in scattering of dielectric surfaces. The novel method takes advantages of the fast decaying property of Green function in lower medium along with the propagating distance. This means that if one excite one point on the surface, only nearby points are being influenced due to the fast decaying property of waves in a lossy medium. The band nature relating fields to each other is utilized and is further found to lead to a fast convergence. Compared to IBC which local electric field and magnetic field are related by a simple relationship of approximation, this new method exploits the band nature of green function and formulates the correct relationship of surface electric field and magnetic fields. We name this method the Neighborhood Impedance Boundary Condition (NIBC). Through our results, it is shown that NIBC is both stable and accurate for different wind speeds and can also be applied to 3D problems. Combined with fast algorithm, NIBC can be both computational fast and efficient. The permittivity used in this chapter are 75+61i except otherwise indicated. Ocean spectrum of Pierson-Moskowitz is used. The chapter is organized as follows: in section 4.2, we are presenting the integral equations used. A tapered incidence wave is also introduced. In 4.3, Pierson-Moskowitz is described. This is followed up with the introduction of NIBC method. Then in the final results and discussion part, first we will show that dual SIE is more accurate than IBC. Then it is proved that dense grid is needed for lossy medium to achieve correct results. Then a slow convergence problem is described when dense grid is implemented compared to coarse grid. 49

66 Finally, we will show that NIBC can alleviate the slow convergence problem but at the same time can be both accurate and stable. 4.2 Formulations In this section, we will introduce the formulations used in IBC and dual SIE. Also tapered wave incidence is described in this section Dual surface integral equation For the 2D rough surface scattering problem, the governing dual surface integral equation (SIE) derived from extinction theorem from both air and medium are [49] 2 dx ' x' 1 f ' x' g x, z; x', z ' g x, z; x', z ' u x' inc x, z n ' (55) z' f x ' 2 1 dx ' x' 1 f ' x' g1 x, z; x', z ' g1 x, z; x', z ' u x' 0 n (56) ' z' f x ' u x 1 f ' x x n 2 (57) where g(x, z; x, z ) is Green function in the air which is i H (1) 4 0 (kρ). k is the wavenumber in the air. g 1 (x, z; x, z ) is Green function in the lower medium which is i H (1) 4 0 (k1 ρ). k 1 is the wavenumber in the lower medium. Ψ is the electric field if it is TE polarization and is magnetic field if it is TM polarization. k k (58) 50

67 f(x) is the rough ocean surface profile and f (x) is the slope of the surface. n is the normal vector of the surface. ψ and u are surface fields. And ρ is an indicator of TE and TM incidence. ψ inc is the incidence wave. Then after performing point matching to the above two equation, we get Au B inc (59) A1u B 1 0 (60) For matrix elements in (59), we have A g x, z ; x, z x for m n (61) mn m m n n i 2 kx A 1 ln 1 ' 2 mm x i f xm 4 4e 2 B x 1 f ' x g x, z ; x', z ' for m n mn m n ' m m (62) (63) 1 Bmm (64) 2 For matrix elements in (60), we have Amn,1 g1 xm, zm; xn, zn x for m n (65) i 2 kx 2 m 4 4e 1 Amm x 1 i ln 1 f ' x 2 B,1 x 1 f ' x g1 x, z ; x', z ' for m n mn m n ' m m (66) (67) 1 Bmm (68) 2 51

68 For usual perfect electric conductor (PEC) scattering problem, 10 points per free space wavelength is usually enough to get correct results. But for dielectric medium with a large permittivity 10 points per wavelength is far from enough. A usual criteria is to discretize the gird so that there are around 10 points per medium wavelength Impedance boundary condition For the assumption of impedance boundary condition, the surface electric field and surface magnetic field is related through the following relationship E Z n H (69) tan s where Z s is the impedance of the lower medium. For two dimensional problem, we have 1 1 n Zs i i 1 u f x 2 for TM polarization (70) 1 1 n Zs i i 1 u f x 2 for TE polarization (71) Thus for either TM or TE polarization we can substitute (70) or (71) into equation (55) to get a simplified equation which only has one kind of unknowns Tapered incidence wave Similar to 3D problems, a tapered incident wave is used in the simulation due to the fact that the surface has to be truncated to a finite length of L. To avoid artificial edge diffraction, the tapered incident wave is defined as 52

69 inc x, z expik xsini z cosi 1 wexp x ztan g 2 i 2 2 (72) where g is the tapering parameter and 2 x ztani g w (73) 2 kg cos i The incident wave is close to a plane wave for large surface length L. The tapering parameter is taken as g = L/ Pierson-Moskowitz spectrum As a function of wavenumber, gravity and wind speed the form of Pierson-Moskowitz spectrum is [14] 2 a g WPM k 4 k 2 4 2k k U19.5 0, exp, (74) where a 0 = , β = 0.74, g is the gravitational acceleration equation to 9.81 m/sec 2. And U19.5 is the wind speed at the altitude of 19.5 m above the mean sea level. In (74), Φ(k, φ) is the directional factor. But since we are exploring the two dimensional problem, Φ(k, φ) is set to be 1 in this chapter. 53

70 Figure 19 Amplitude of the Pierson-Moskowitz spectrum for along-wind direction and wind speeds U19.5=4, 8 and 10 m/s Figure 19 plots the amplitude of the Pierson-Moskowitz spectrum for along-wind direction and wind speeds U19.5=4, 8 and 10 m/s. It is seen that much of the energy is concentrated on small wavenumber. And with an increase of wind speed, the peak of the spectrum is moving towards the small wavenumber. This means that longer ocean waves are included in higher wind speeds. 4.4 Neighborhood impedance boundary condition In this section we are describing the proposed novel method for scattering of highly dielectric lossy surfaces. Different from usual method which solves (59) and (60) directly, the proposed NIBC method is trying to solve (60) first. We then have u A1B 1 M (75) M is the matrix relating u and ψ which is in contrast to the relationship defined in (70) and (71).Substitute (75) into (59), gives 1 AA1 B1 B inc (76) 54

71 (76) then becomes the final equation we are solving. In the above procedure, we are facing the following 3 issues. A. Inversion of matrix A 1 As mentioned earlier, since the permittivity of sea water has a large imaginary part, the waves will decay fast with the travelling distance. Thus the matrix A 1 and B 1are actually band matrix. We can set up a bandwidth BW of the matrices that the matrix entries outside the BW is set to be zero. Figure 20 shows a comparison of 3D green function of air and sea water in terms of absolute value. It is easy to see that the green function in the sea water decays very fast. Even for a propagating distance of one λ, the absolute value drops to which is much lower than the green function in the air. This means that A 1 and B 1 actually behaves like a band matrix. The matrix elements quickly decrease when moving away from the diagonal line. Figure 20 3D Green function comparison of air and sea water as a function of propagating 55

72 In our numerical implementation, the BW is configured that the furthest elements in the matrix are 10-4 lower than the diagonal part. It is observed that when BW equal 1 λ can meet such 4 a requirement. In the results section, we will show that the accuracy varies little with the BW used in the simulation. Now the problems reduces to solve for the inverse matrix of a diagonal matrix with a bandwidth BW. The following is an expression of the definition of inversion. 1 A1A1 I (77) In order to obtain the inverse of A 1, we can solve for each column of A 1 1 one at a time. That is each time we are solving A1a n e n (78) where a n is the nth column of matrix A 1 and e n is a column vector with all zero elements except the nth element. Figure 21 Fast method for band matrix inversion 56

73 Figure 21 shows an example about how we invert the matrix. In this case, the BW is 3 grids and the figure is showing a case to solve for the 4 th column of the matrix A 1 1. Then we just need to solve an inversion problem of size 5 by 5. The x and d in the figure represent diagonal and off-diagonal elements, respectively. Thus compared to usual method of inversing a matrix which is of the computation complexity order O(N 3 ), our method will only require O(N BW 3 ). B. Computation of A 1 1 B 1b A 1 1 and B 1are both band matrix with bandwidth around BW. We can do first step B1b c (79) The computational complexity is of order O(N BW). Then 1 A1 c d (80) The computational complexity is also of order O(N BW). Thus a nice feather of band matrix is that it is not so computation expensive while conducting matrix vector multiplication. C. A b and B b We have implemented PBTG/SMCG to accelerate this part. The details are described in the next chapter. When doing matrix times vector, since B 1, A 1 1 are set to be band matrix, only entries within the bandwidth need to be calculated and stored. This has in fact reduced the memory storage. 4.5 Results and Discussions There are 3 methods in this section: IBC, NIBC, and benchmark solutions. Benchmark solutions refer to the results by directly solving equations (59) and (60). All the cases in this section 57

74 are TM incidence cases except otherwise indicated. The dielectric constant is 75+61i and the MoM is conducted on a mesh with 64 points per free space wavelength. The ocean spectrum of Pierson Moskowitz spectrum is used. The error of using Frobenius norm is used to calculate the difference between two arrays. It is defined as array1 array2 F error (81) array1 F where F is the Frobenius norm C n1 F n i1 i 2 c (82) Comparison between IBC and benchmark solution Figure 22 Surface fields comparisons between IBC and benchmark solution for TE incidence at 5m/s. 58